See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/339984770

What We Talk About When We Talk About Ped-Bike Collisions in YouTube

Preprint · May 2020

CITATIONS READS
0 53

3 authors:

Subasish Das Xiaoqiang Kong

Texas A&M University Texas A&M University
85 PUBLICATIONS   311 CITATIONS    5 PUBLICATIONS   11 CITATIONS   

SEE PROFILE SEE PROFILE

Ruihong Wang
Texas A&M University Central Texas
2 PUBLICATIONS   0 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Transportation Safety Planning (TSP) Noteworthy Practices View project

Coauthor Network Project View project

All content following this page was uploaded by Subasish Das on 20 March 2020.

The user has requested enhancement of the downloaded file.

 What We Talk About When We Talk About
Ped-Bike Collisions in YouTube
Subasish Das Xiaoqiang Kong
Texas A&M Transportation Institute Texas A&M Transportation Institute
Bryan, TX- 77807 Bryan, TX- 77807
s-das@tti.tamu.edu x-kong@tti.tamu.edu
Abstract—Many studies have examined vehicle-pedestrian and
vehicle-bicycle crash patterns. However, there is a surprisingly
limited amount of research focused on pedestrian-bicyclist colli-
sions, and research about the tension between pedestrians and
cyclists is even rarer. The lack of research on this subject is partly
due to the limited number of pedestrian-cyclist crashes; it is also
due to the fact that the consequences of these crashes are typically
less severe than those of vehicle-involved crashes. Despite the lack
of research focus on this area, pedestrian-cyclist crashes could
lead to a serious social crisis. The largest video sharing website,
YouTube.com, contains many pedestrian-cyclist collision videos
and associated comments and replies. The application of content
analysis and text mining to the comments from these videos
can provide insight into potential interactions. The findings of
this study show that the emotion patterns of comments and
replies differ. This study also provides word shift plots that
show the trend of the emotion used in comments and replies.
Additionally, the co-occurrence plots show the reason behind
the use of negative emotions. The findings of this study will
provide additional insights into the ongoing debate on ‘pedestrian
collisions with bicyclists’ issues.
Index Terms—autonomous vehicles, content analysis, text min-
ing, sentiment analysis.
I. I NTRODUCTION
In recent years, non-motorized travel modes (walking and
biking) have been gaining popularity in the U.S. with non-
motorist travelers. To improve the safety of pedestrians and
bicyclists, researchers have made multiple efforts to reduce
vehicle-bicycle and vehicle-pedestrian crashes. A research area
that is less explored is the collision between a bicyclist
and pedestrian. A recent study examined the incidences of
pedestrians injured by cyclists in California from 2005 to
2011 and in New York between 2004 to 2011 [34]. The
findings show that despite the increasing number of cyclists,
there was a decline in the rate of pedestrians injured in
collisions with cyclists. One of the primary reasons for the rate
decrease is due to the cycling infrastructure improvements.
In the absence of cycling infrastructures, cyclists often use
pedestrian facilities for part or the entire journey. Pedestrians
feel insecure in the presence of speedy bicyclists, especially
in dense urban locations. In recent years, antagonism between
pedestrians and bicyclists has gained more attention through
social media like mainstream media websites. It is easy to
find videos and articles that contain furious comments about
careless pedestrians or bicyclists online. The tensions between
Ruihong Wang
Texas A&M Transportation Institute
Bryan, TX- 77807
r-wang@tti.tamu.edu
pedestrians and bicyclists are hard to resolve because both are
vulnerable road users. Pedestrians may seem more vulnerable
to most of the public, but this belief leads to a violent,
negative image of bicyclists [1]. Mainstream websites present
contradicting arguments between the bicyclists and the rest of
the public. However, the lack of related academic studies will
bridge the gap by exploring the causes of tensions between
bicyclists and pedestrians by performing topic modeling and
text mining.
In recent years, the role that social media can have in
shaping the opinions of individuals on various products and
issues has gained attention. Because videos allow the visu-
alization of information, concepts, and dialogues and permit
user-generated communications, videos have developed public
perceptions. YouTube.com has more than 1 billion users and is
the largest online platform for open-access video content [24].
As such, this platform plays a significant role in generating
public opinion on many issues. YouTube contains the highest
number of public opinions on video relevant interactions in
comparison to other social media platforms.
To perform knowledge discovery on the motives of user
participation and consumption on YouTube videos associated
with the ‘pedestrian collisions with bicyclists,’ this study
applies the natural language processing (NLP) framework.
This study aims to explore to research questions: 1) R1:
What are the key patterns and trends of bicyclist vs pedestrian
interactions? and 2) R2: Do comments and replies vary in
sentiment scores by video types?
II. EARLIER WORK AND RESEARCH CONTEXT
Very little research has concentrated on the collision be-
tween cyclists and pedestrians. Moreover, any mention of
tension between pedestrians and cyclists in research is rare
because of the limited number of crashes; the exception is in
some urban-core ped-bike active zones. More serious issues
could be a result of the tension between the two groups [2],
[34].
Tuckel et al. conducted a study to examine the trend
of pedestrian injuries in the bicyclist-pedestrian (bike-ped)
collisions and investigate possible explanations [34]. The New
York Times published an article called “The Cyclist-Pedestrian
Wars,” detailing the rising tensions between bicyclists and
pedestrians in the Upper East Side of New York [15]. An
 Fig. 1: Flowchart of data collection and analysis.

TABLE I: Clusters of Top-viewed 10 YouTube Videos on ‘Bicyclists Hitting Pedestrians’

Publish Duration
Video Id Title Views Likes Dislikes Comments
Date (min.)
zR4Okh23Zlo Cyclist hits pedestrian 11-Aug-13 2:06 3,099,255 31,000 2,700 14,737
Pedestrian gets hit,
sYWPHHo0fPU 26-Oct-16 6:33 2,255,886 12,000 4,300 7,939
cyclists talk, cops
Cyclist crashes into
Wq6rpVMcyas 12-Nov-17 0:59 748,883 8,300 187 1,424
man full video
Angry pedestrian blocks
G4K8AjNIVPA cyclist as he races 28-Sep-16 0:27 401,683 2,500 146 1,429
through zebra crossing
0Lm9TPym9A4 Man gets hit by bike 14-Aug-10 0:37 250,312 1,500 288 371
Pedestrian and bicycle
5Qurlf05YYI 1-Apr-13 0:39 20,050 58 15 77
accident on Venice Beach
Pedestrian hit by
dnkErN9N8KY 15-Mar-17 1:44 9,989 51 7 41
bicycle in San Francisco
uRoU826ywjw Cyclist hits pedestrian 18-May-15 0:44 6,138 25 23 55
Accident on the
dXpmxmFW164 6-Apr-15 1:50 4,585 26 6‘ 23
bicycle lane
Pedestrian almost
s-PuD8fSI-I 15-Jul-14 0:34 3,157 22 4 26
hit by cyclist

Amsterdam study showed that traffic conflicts with bicycle
paths are a specific safety problem for crossing pedestrians
[35]. The bicyclist blames the pedestrian for stepping into
the bike lane illegally, and the pedestrian blames the bicyclist
for running too quickly and without looking [30]. In Santa
Monica, California, the tension between the pedestrians and
bikers was explained in a published news article [19]. Werneke
et al. found that pedestrians crossing the bicycle path without
looking were the primary cause of the majority of bike-
pedestrian incidents [36].
In the recent years, several studies have incorporated text
mining in transportation engineering research: consumer com-
plaint analysis [14], [21], social media mining [7], [13],
[16], [28], opinion mining on safety enhancement and bike-
sharing [5], [10], topic modeling on transportation engineering
conference papers and journals [3], [6], [9], [17], [33] and
crash narrative investigation [4], [8]. An investigation into
the textual content related to YouTube videos on ‘pedestrian
collisions with bicyclists’ has not yet been conducted. A debate
on whether social media data is representative and unbiased
enough for a robust study. This study contemplates if the
collection of 25,000 comments and replies will provide an
accurate representation of public opinion and sentiments on
this issue.
III. DATA DESCRIPTION
A. Data Collection
To collect the ‘bicycle hitting pedestrian’ related videos
in YouTube, a detailed list of keywords was developed by
using the following terms: “walking biking collision,” “biker
hits ped,” “bicyclist hit pedestrian,” “pedestrian bike crash or
incident or accident,” “pedestrian bicyclist crash or incident
or accident.” The researchers automated the data collection
(extract the video information as well as related comments)
process by using an open-source R software package called
“tuber” [32]. Another online YouTube comment scrapper has
also been used [20]. The research team used several open-
source R software packages to perform the analysis [22], [29],
[31], [37]. The flowchart of data collection and analysis is
shown in Fig. 1.
Table I provides descriptive statistics of the top ten most
viewed videos. After removing redundant and non-English
comments, the final number of comments was 26,122. Among
the top ten videos, one video was released earlier (in 2010).
 (a) Comment
Fig. 2: Valence shift word graphs based on comments and replies.
The overall number of views for all videos was 6,799,938
(mean: 679,994, standard deviation: 1,098,276). The number
of likes on all videos was higher than dislikes (55,482 vs.
7,670). The number of comments were 26,122 (mean: 2,612;
standard deviation: 4,901). In these videos, participants also
replied to the comments. The replies to the comments are also
collected and analyzed in this study. The corpora have around
2,000 replies, based on all replies.
IV. METHODOLOGY
Natural language processing (NLP), a popular branch of
computer science, tends to view the process of text and
language analysis as being divided into a number of stages by
conducting theoretical linguistic distinctions between syntax
or parsing (relationship between words), semantics (meaning
of a textual content), and pragmatics (process of extracting
information from text). Typically, sentences retrieved from a
text or document are first analyzed in terms of the associated
syntax, which provides a procedure that is more suitable
to an analysis in terms of semantics and other associated
meaning [18]. All required NLP tasks such as tokenization,
lemmatization, parts of speed tagging, and dependency parsing
were conducted before starting the next steps.
A. Term Frequency-Inverse Document Frequency
Instead of using a word or word group frequencies, another
approach is to look at a term’s inverse document frequency
(b) Reply
(IDF). Spark Jones first introduced this concept in 1972 [31].
Table II shows the TF-IDF values for the two categories based
on interaction types. All comments or replies for each of
these videos are combined by video ids for determining TF-
IDF measures. As unigrams are not suitable in explaining the
intent of the topics, bigrams are considered in this analysis.
A threshold of 200 counts is considered as the baseline for
comment corpora. For the reply corpora, this threshold was
20. The majority of the bigrams overlap in both categories.
Intersection, signal phases, bike lanes, and lighting conditions
are the most common bigrams in both categories. The bigrams
‘the crosswalk’ and ‘parents fault’ are present in the comment
category analysis. In the reply categories, two unique bigrams
are ‘walk on’ and ‘walk in.’
B. Sentiment Analysis
Mining on subjective texts containing opinion or sentiment
can contribute to understanding perception towards a product.
In other words, the objective of sentiment analysis is to de-
termine which words or sentences express opinions, feelings,
and sentiments. The sentiment score can be easily calculated
by using the number of positive words or sentences minus the
number of negative words or sentences. The research team
used ‘udpipe’ inbuilt functions to determine the sentiment
scores [37]. The descriptive statistics of the sentiment scores
by the video ids are shown in Table III. Each video id
is listed with the maximum, minimum, mean, and standard
 TABLE II: TF-IDF of the Top Bigrams from Comment and Reply Corpora
VID Bigram TF IDF TF-IDF
Comments
sYWPHHo0fPU the light 0.00331 1.20397 0.00398
zR4Okh23Zlo bike lane 0.00567 0.69315 0.00393
sYWPHHo0fPU the intersection 0.00197 1.60944 0.00317
sYWPHHo0fPU yellow light 0.00124 2.30259 0.00286
sYWPHHo0fPU red light 0.00232 1.20397 0.00279
sYWPHHo0fPU light was 0.00152 1.60944 0.00245
zR4Okh23Zlo bike path 0.00218 0.91629 0.002
zR4Okh23Zlo parents fault 0.0007 2.30259 0.00162
sYWPHHo0fPU the crosswalk 0.00146 0.91629 0.00134
sYWPHHo0fPU slow down 0.00177 0.69315 0.00123
Replies
sYWPHHo0fPU the light 0.004216 1.504077 0.006342
zR4Okh23Zlo walk on 0.001681 2.197225 0.003694
zR4Okh23Zlo bike path 0.001639 2.197225 0.003602
sYWPHHo0fPU the intersection 0.00233 1.504077 0.003505
sYWPHHo0fPU light was 0.001498 2.197225 0.003291
sYWPHHo0fPU was red 0.001387 2.197225 0.003047
zR4Okh23Zlo cycle lane 0.001303 2.197225 0.002863
sYWPHHo0fPU red light 0.002441 1.098612 0.002682
zR4Okh23Zlo walk in 0.001177 2.197225 0.002586
sYWPHHo0fPU slow down 0.002164 1.098612 0.002377

deviation of each comment and reply. The video with the
highest comment average is dnkErN9N8KY (Pedestrian hit
by bicycle in San Francisco) with 0.39. It also has the highest
maximum, minimum, and standard deviation measures.
1) Valence Shift Word Graphs: In their study, Dodds and
Danforth [11] provided the importance of ‘Valence Shift
Word Graph.’ Consider two texts Tref (for reference) and
(ref )
Tcomp (for comparison) with sentiment scores smean and
(comp)
smean . A word’s sentiment is signified relative to text Tref
by + (positive sentiment) and - (negative sentiment), and its
relative abundance in text Tcomp versus text Tref with ↑ (more
prevalent) and ↓ (less prevalent). Combining these two binary
possibilities
P leads to four cases [12].
Where i δsmean,i = ±100, depending on the sign of the
(comp) (ref )
difference in sentiment between the two texts, smean −smean
, and the terms to which the symbols +/− and ↑↓ apply
have been indicated. The δsmean,i is called the per word
sentiment shift of the ith word. Figure 4 can be interpreted
by the following interpretations:
• Words on the right contribute to an increase in positive
emotions in the corpus
• A yellow bar in the right with a down arrow indicates
that a negative emotion was used less
• A purple bar in the right with an up arrow indicates that
positive emotion was used more
• Words on the left contribute to a decrease in position
emotions in the corpus
• A yellow bar in the left with an up arrow indicates that
a negative emotion was used more
• A purple in the left with a down arrow indicates that
positive emotion was used less
The word shift plots are not significantly different between
the corpora (plural of corpus) developed for comments and
replies. However, the degree of negative emotions is less
used in the replies. Some of the terms, such as bike, are
considered as positive emotion due to the reason of using
conventional sentiment lexicons. There is a need for the
development of transportation safety-related senti-lexicon to
precisely capture the domain-specific sentiments and emotions,
which is currently out of the scope of the present study.
C. Emotion Mining
Emotion mining, a similar method like sentiment analy-
sis, detects, analyzes, and performs evaluations on humans’
feelings towards different issues, topics, and scenarios. A
specific direction of emotion mining includes text emotion
mining, which refers to the examination of people’s emotions
based on their writing observations. By manual annotation
through Amazon’s Mechanical Turk service, Mohammad and
Turney [23] compiled a large English term–emotion associ-
ation lexicon, named EmoLex. The lexicon focused on the
emotions of joy, anger, sadness, fear, disgust, trust, surprise,
and anticipation and has been argued by many to be the
prototypical and basic emotions [25]–[27].
For the emotion mining tasks, this study considers eight
major emotion types and their negations. The trends of the
emotions are shown at the sentence level (shown in Fig. 3).
This method uses sentiment lexicons to find emotion words
and then compute the emotion propensity per sentence [37].
The x-axis indicates the number of documents in percentage
form. For example, if the analysis is conducted on 100
documents, 25 percent will indicate the 25th document, and if
the vertical line is drawn on 25%, the intersecting points will
be the emotion propensity score for that particular sentence.
D. Co-occurrence of Negative Terms
The majority of the sentiment analysis and emotion mining
studies perform only n-gram related studies to determine the
 TABLE III: Descriptive Statistics of Sentiment Scores by Videos
Max Min Mean STD
VID
Comment Reply Comment Reply STD Reply Comment Reply
0Lm9TPym9A4 4.10 0.80 -6.00 -2.65 -0.43 -0.71 1.22 0.88
5Qurlf05YYI 2.25 2.80 -3.25 -0.75 -0.41 0.75 1.01 1.20
dnkErN9N8KY 5.85 5.85 -1.50 -2.50 0.39 -0.03 1.53 1.72
dXpmxmFW164 1.05 0.80 -1.00 -1.00 0.26 -0.07 0.75 0.90
G4K8AjNIVPA 3.10 3.10 -8.60 -4.85 -0.28 -0.30 1.13 1.22
s-PuD8fSI-I 1.40 1.40 -1.75 -1.40 -0.13 -0.04 0.84 0.89
sYWPHHo0fPU 7.60 – -7.15 – -0.37 – 1.06 –
uRoU826ywjw 2.80 4.80 -0.75 -7.15 0.68 -0.32 1.15 1.27
Wq6rpVMcyas 3.00 1.00 -4.50 -4.10 -0.33 -0.38 0.85 0.84
zR4Okh23Zlo 5.30 5.30 -10.00 -6.45 -0.49 -0.42 1.15 1.32

co-occurrence. The words’ co-occurrence relationships for the

datasets of comments and replies are shown in Fig. 4. The
general trend is that negative sentiment patterns are not same
in comment corpus and reply corpus.

(a) Emotions from Main Comments

(a) Words links with negative terms (Dataset of Comments)

(b) Emotions from Replies of Comments

Fig. 3: Emotion patterns.

sentiments and emotions over the corpus, document, or sen-

tence level. One area that is less explored is the investigation
on determining the relation of other words with the negative
sentiments and emotions. This approach will help answer what
is causing a negative sentiment or negative emotion. The
Stanford Dependencies (SD) representation was originally de-
veloped as a practical representation of English syntax, aimed
at natural language understanding (NLU) applications. This
is deeply associated with grammatical relation-based syntactic (b) Words links with negative terms (Dataset of Replies)
concepts. This study used the dependency relationship output
of ‘udpipe’ to find out which words are linked to negative Fig. 4: Cooccurences of the negative terms.
words from the ‘udpipe’ sentiment dictionary [37]. Out of
several parameters, this study used mainly the parameters
V. CONCLUSIONS
associated with adjectives that modify a noun. Before conduct-
ing the dependency parsing, this study used the conventional The understanding of the public perception of ‘bicyclist
NLP annotation (tokenization, lemmatization, parts of speech versus pedestrian collision’ is important. Due to the lack
tagging). The lemma values of the negative words and the of adequate and quick information collection and retrieval,
lemma values of the parent word are used to calculate the conventional surveys are not sufficient in comprehending the
tipping point of consumer perception. By performing an anal-
ysis of responses to pedestrians and cyclist’s collision videos
on YouTube, this study advances the general understanding of
the communication between the bicyclists and pedestrians. The
study concentrates on the content exploration of the YouTube
videos by evaluating the behaviors of the end-users regarding
disliking, liking, and commenting patterns.
The findings of this study answer the key research questions.
Some of the key findings are:
• A large number of comments, views, likes, and dislikes
can indicate that the public is partaking in this debate.
• The tf-idf algorithm identifies multiple rare but significant
words at various categorization levels.
• The valence shift word graphs provide a synopsis of the
contexts of the words and the shifting of emotions in the
replies and comments. The outcomes show that reply and
comment patterns vary in sentiment measures.
• The co-occurrence plots show the logic behind the nega-
tive emotions generated in the replies and comments.
The findings from this study can benefit the development
of better communication between bicyclists and pedestrians.
Additionally, this study developed a framework to determine
consumer’s adaptation needs that can be replicated for other
transportation-related topics. Although this study contributed
positively to research, this current study has several limi-
tations. One of the critical limitations is the small sample
size. This study only evaluated the top ten viewed videos and
related comments. Future studies should expand data or collect
data from other video platforms like Facebook or Vimeo to
understand the interactions between pedestrians and bicyclists.
R EFERENCES
[1] Isolatecyclist. cyclists versus pedestrians, 2012.
[2] Soufiane Boufous, Julie Hatfield, and Raphael Grzebieta. The impact
of environmental factors on cycling speed on shared paths. Accident
Analysis & Prevention, 110:171–176, 2018.
[3] Ryan C Boyer, William T Scherer, and Michael C Smith. Trends over
two decades of transportation research: a machine learning approach.
Transportation research record, 2614(1):1–9, 2017.
[4] Donald E Brown. Text mining the contributors to rail accidents.
IEEE Transactions on Intelligent Transportation Systems, 17(2):346–
355, 2015.
[5] Feng Chen, Ramayya Krishnan, et al. Transportation sentiment analysis
for safety enhancement. 2013.
[6] S Das and A Dutta. Knowledge extraction from transportation research
thesaurus. In 97th Annual Meeting of the Transportation Research
Board, Washington, DC, 2018.
[7] Subasish Das. Social media interactions from transportation’s largest
conference. TR NEWS, 2019.
[8] Subasish Das, Bradford K Brimley, Tomas E Lindheimer, Ashesh Pant,
et al. Safety impacts of reduced visibility in inclement weather.
Technical report, Center for Advancing Transportation Leadership and
Safety (ATLAS Center), 2017.
[9] Subasish Das, Karen Dixon, Xiaoduan Sun, Anandi Dutta, and Michelle
Zupancich. Trends in transportation research: Exploring content analysis
in topics. Transportation Research Record, 2614(1):27–38, 2017.
[10] Subasish Das, Xiaoduan Sun, Anandi Dutta, et al. Investigating user rid-
ership sentiments for bike sharing programs. Journal of Transportation
Technologies, 5(02):69, 2015.
[11] Peter Sheridan Dodds and Christopher M Danforth. Measuring the
happiness of large-scale written expression: Songs, blogs, and presidents.
Journal of happiness studies, 11(4):441–456, 2010.
View publication stats
[12] Peter Sheridan Dodds, Kameron Decker Harris, Isabel M Kloumann,
Catherine A Bliss, and Christopher M Danforth. Temporal patterns of
happiness and information in a global social network: Hedonometrics
and twitter. PloS one, 6(12), 2011.
[13] Tamer El-Diraby, Amer Shalaby, and Moein Hosseini. Linking social,
semantic and sentiment analyses to support modeling transit customers’
satisfaction: Towards formal study of opinion dynamics. Sustainable
Cities and Society, 49:101578, 2019.
[14] Mahtab Ghazizadeh, Anthony D McDonald, and John D Lee. Text
mining to decipher free-response consumer complaints: Insights from the
nhtsa vehicle owner’s complaint database. Human factors, 56(6):1189–
1203, 2014.
[15] J. Goodman and D. Spokes. The cyclist-pedestrian wars. the new york
times, 2010.
[16] Yiming Gu, Zhen Sean Qian, and Feng Chen. From twitter to detector:
Real-time traffic incident detection using social media data. Transporta-
tion research part C: emerging technologies, 67:321–342, 2016.
[17] Jungyeol Hong, Reuben Tamakloe, Gunwoo Lee, and Dongjoo Park. In-
sight from scientific study in logistics using text mining. Transportation
research record, 2673(4):97–107, 2019.
[18] Aleš Horák, Ivan Kopecek, and Karel Pala. Text, speech and dialogue.
Springer, 2010.
[19] G. Kavanagh. Unraveling ped & bike tension in santa monica, 2012.
[20] P. Klostermann. A web client that scrapes youtube comments.
[21] S Mehrotra and S Roberts. Identification and validation of themes from
vehicle owner complaints and fatality reports using text analysis. In 97th
Annual Meeting of the Transportation Research Board, Washington, DC,
2018.
[22] David Meyer, Kurt Hornik, and Ingo Feinerer. Text mining infrastructure
in r. Journal of statistical software, 25(5):1–54, 2008.
[23] Saif M Mohammad and Peter D Turney. Crowdsourcing a word–emotion
association lexicon. Computational Intelligence, 29(3):436–465, 2013.
[24] John A Naslund, Stuart W Grande, Kelly A Aschbrenner, and Glyn
Elwyn. Naturally occurring peer support through social media: the
experiences of individuals with severe mental illness using youtube.
PLOS one, 9(10), 2014.
[25] Robert Plutchik. A general psychoevolutionary theory of emotion. In
Theories of emotion, pages 3–33. Elsevier, 1980.
[26] Robert Plutchik. On emotion: The chicken-and-egg problem revisited.
1985.
[27] Robert Plutchik. The psychology and biology of emotion. HarperCollins
College Publishers, 1994.
[28] Ali Rahim Taleqani, Jill Hough, and Kendall E Nygard. Public opinion
on dockless bike sharing: A machine learning approach. Transportation
research record, 2673(4):195–204, 2019.
[29] T. Rinker. sentimentr: Calculate text polarity sentiment.
[30] F. Salmon. A unified theory of new york biking. reuters, 2010.
[31] Julia Silge and David Robinson. Text mining with R: A tidy approach.
” O’Reilly Media, Inc.”, 2017.
[32] G Sood. tuber: Access youtube from r. r package version 0.9. 7, 2018.
[33] Lijun Sun and Yafeng Yin. Discovering themes and trends in trans-
portation research using topic modeling. Transportation Research Part
C: Emerging Technologies, 77:49–66, 2017.
[34] Peter Tuckel, William Milczarski, and Richard Maisel. Pedestrian
injuries due to collisions with bicycles in new york and california.
Journal of safety research, 51:7–13, 2014.
[35] A Richard A van der Horst, Maartje de Goede, Stefanie de Hair-
Buijssen, and Rob Methorst. Traffic conflicts on bicycle paths: A
systematic observation of behaviour from video. Accident Analysis &
Prevention, 62:358–368, 2014.
[36] Julia Werneke, Marco Dozza, and MariAnne Karlsson. Safety–critical
events in everyday cycling–interviews with bicyclists and video annota-
tion of safety–critical events in a naturalistic cycling study. Transporta-
tion Research Part F: Traffic Psychology and Behaviour, 35:199–212,
2015.
[37] Jan Wijffels, Milan Straka, and Jana Strakov. udpipe: Tokenization,
parts of speech tagging, lemmatization and dependency parsing with
the’udpipe”nlp’toolkit. R package version 0.5, 2018.